1 # KCF tracker – parallel and PREM implementations
3 The goal of this project is modify KCF tracker for use in the
4 [HERCULES][1] project, where it will run on NVIDIA TX2 board. To
5 achieve the needed performance we try various ways of parallelization
6 of the algorithm including execution on the GPU. The aim is also to
7 modify the code according to the PRedictable Execution Model (PREM).
9 Stable version of the tracker is available from [CTU server][2],
10 development happens at [GitHub][3].
12 [1]: http://hercules2020.eu/
13 [2]: http://rtime.felk.cvut.cz/gitweb/hercules2020/kcf.git
14 [3]: https://github.com/Shanigen/kcf
18 The code depends on OpenCV 2.4 library
19 and [CMake][13] (optionally with [Ninja][8]) is used for building.
20 Depending on the version to be compiled you need to have development
21 packages for [FFTW][4], [CUDA][5] or [OpenMP][6] installed.
23 On TX2, the following command should install what's needed:
25 $ apt install cmake ninja-build libopencv-dev libfftw3-dev
28 [4]: http://www.fftw.org/
29 [5]: https://developer.nvidia.com/cuda-downloads
30 [6]: http://www.openmp.org/
31 [13]: https://cmake.org/
35 There are multiple ways how to compile the code.
37 ### Compile all supported versions
40 $ git submodule update --init
44 This will create several `build-*` directories and compile different
45 versions in them. If prerequisites of some builds are missing, the
46 `-k` option ensures that the errors are ignored. This uses [Ninja][8]
47 build system, which is useful when building naively on TX2, because
48 builds with `ninja` are faster (better parallelized) than with `make`.
50 To build only a specific version run `make <version>`. For example,
51 CUDA-based version can be compiled with:
57 [8]: https://ninja-build.org/
62 $ git submodule update --init
67 - Use the just created build directory as "Where to build the
70 - Choose desired build options. Each option has a comment briefly
71 explaining what it does.
72 - Press "Generate" and close the window.
80 $ git submodule update --init
83 $ cmake [options] .. # see the tables below
87 The `cmake` options below allow to select, which version to build.
89 The following table shows how to configure different FFT
92 |Option| Description |
94 | `-DFFT=OpenCV` | Use OpenCV to calculate FFT.|
95 | `-DFFT=fftw` | Use fftw and its `plan_many` and "New-array execute" functions. If `std::async`, OpenMP or cuFFTW is not used the plans will use 2 threads by default.|
96 | `-DFFT=cuFFTW` | Use cuFFTW interface to cuFFT library.|
97 | `-DFFT=cuFFT` | Use cuFFT. This version also uses pure CUDA implementation of `ComplexMat` class and Gaussian correlation.|
99 With all of these FFT version additional options can be added:
101 |Option| Description |
103 | `-DASYNC=ON` | Use C++ `std::async` to run computations for different scales in parallel. This doesn't work with `BIG_BATCH` mode.|
104 | `-DBIG_BATCH=ON` | Concatenate matrices of different scales to one big matrix and perform all computations on this matrix. This mode doesn't work with `OpenCV` FFT.|
105 | `-DOPENMP=ON` | Parallelize certain operation with OpenMP. This can only be used with `OpenCV` or `fftw` FFT implementations. By default it runs computations for differenct scales in parallel. With `-DBIG_BATCH=ON` it parallelizes the feature extraction and the search for maximal response for differenct scales. With `fftw`, Ffftw's plans will execute in parallel.|
106 | `-DCUDA_DEBUG=ON` | Adds calls cudaDeviceSynchronize after every CUDA function and kernel call.|
107 | `-DOpenCV_DIR=/opt/opencv-3.3/share/OpenCV` | Compile against a custom OpenCV version. |
110 ### Compilation for non-TX2 CUDA
112 The CuFFT version is set up to run on NVIDIA Jetson TX2. If you want
113 to run it on different architecture, change the `--gpu-architecture
114 sm_62` NVCC flag in **/src/CMakeLists.txt** to your architecture of
115 NVIDIA GPU. To find what SM variation you architecture has look
118 [9]: http://arnon.dk/matching-sm-architectures-arch-and-gencode-for-various-nvidia-cards/
122 No matter which method is used to compile the code, the results will
125 It operates on an image sequence created according to [VOT 2014
126 methodology][10]. You can find some image sequences in [vot2016
129 The binary can be run as follows:
131 1. `./kcf_vot [options]`
133 The program looks for `groundtruth.txt` or `region.txt` and
134 `images.txt` files in current directory.
136 - `images.txt` contains a list of images to process, each on a
138 - `groundtruth.txt` contains the correct location of the tracked
139 object in each image as four corner points listed clockwise
140 starting from bottom left corner. Only the first line from this
142 - `region.txt` is an alternative way of specifying the location of
143 the object to track via its bounding box (top_left_x, top_left_y,
144 width, height) in the first frame.
146 2. `./kcf_vot [options] <directory>`
148 Looks for `groundtruth.txt` or `region.txt` and `images.txt` files
149 in the given `directory`.
151 3. `./kcf_vot [options] <path/to/region.txt or groundtruth.txt> <path/to/images.txt> [path/to/output.txt]`
153 By default the program generates file `output.txt` containing the
154 bounding boxes of the tracked object in the format "top_left_x,
155 top_left_y, width, height".
157 [10]: http://www.votchallenge.net/
158 [11]: http://www.votchallenge.net/vot2016/dataset.html
162 | Options | Description |
163 | ------- | ----------- |
164 | --visualize, -v[delay_ms] | Visualize the output, optionally with specified delay. If the delay is 0 the program will wait for a key press. |
165 | --output, -o <output.txt> | Specify name of output file. |
166 | --debug, -d | Generate debug output. |
167 | --fit, -f[W[xH]] | Specifies the dimension to which the extracted patch should be scaled. It should be divisible by 4. No dimension is the same as `128x128`, a single dimension `W` will result in patch size of `W`×`W`. |
171 * Vít Karafiát, Michal Sojka
173 Original C++ implementation of KCF tracker was written by Tomas Vojir
174 [here][12] and is reimplementation of algorithm presented in
175 "High-Speed Tracking with Kernelized Correlation Filters" paper [1].
177 [12]: https://github.com/vojirt/kcf/blob/master/README.md
181 [1] João F. Henriques, Rui Caseiro, Pedro Martins, Jorge Batista,
182 “High-Speed Tracking with Kernelized Correlation Filters“, IEEE
183 Transactions on Pattern Analysis and Machine Intelligence, 2015
187 Copyright (c) 2014, Tomáš Vojíř\
188 Copyright (c) 2018, Vít Karafiát\
189 Copyright (c) 2018, Michal Sojka
191 Permission to use, copy, modify, and distribute this software for research
192 purposes is hereby granted, provided that the above copyright notice and
193 this permission notice appear in all copies.
195 THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
196 WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
197 MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
198 ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
199 WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
200 ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
201 OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.